The default parameters could not improve after several epochs, resulting in similar images among the later hundreds of epochs. I changed the batch size to 128, discriminator learning rate to 0.0003, and generator learning rate to 0.0001. In this situation, generator can slowly learn after the stronger the discriminator has grown. In the plot, we can observe that the generator loss decreases after about 125 epochs. I assume that the model can improve more if I set a larger number of epochs for training. Codes were changed in WGTrain and WDTrain function.